Exposing the Shared Cache Behavior of Helper Thread on CMP Platforms

Irregular and data-intensive applications are characterized by non-uniform data access patterns and poor temporal and spatial locality, but they can exhibit algorithm locality. Helper thread is a promising prefetching technique to construct the algorithm locality on CMP platforms. However, previous works about helper thread on commercial CMP platforms were oblivious of the relations between the dynamic characteristics of the application´s memory behavior and the parameters in helper thread. Having the insight of dynamic behavior of helper thread, in this paper, we aim to capture the inherent feature under helper thread. We present the HCS (helper thread contribution sensitivity) metric to characterize the dynamic shared cache behavior in helper thread. By using the selected data intensive benchmarks which come from SPEC2006 and Olden suite, the metrics were evaluated on modern commercial CMP platforms. The evaluation results are helpful for optimizing helper thread´s parameters.